are AI detectors accurate?

the honest answer is that it depends on the content and the test, and no single percentage covers it. the largest peer-reviewed benchmark, RAID (ACL 2024), evaluated a dozen detectors across more than six million generations and found that tools advertising 99% or higher accuracy “are easily fooled by adversarial attacks, variations in sampling strategies, repetition penalties, and unseen generative models.” held to a fixed low false-positive rate, even the strongest systems land somewhere around 73-85% at best, and that figure slides further once the text is deliberately altered.

accuracy is meaningless without a false-positive rate

a detection accuracy number only means something next to the false-positive rate it was measured at. the same tool can look excellent at a 5% false-positive rate and close to useless at 0.5%, because tightening the threshold to avoid wrongly accusing humans also lets more AI text slip through. RAID reports its results at a fixed false-positive rate for exactly this reason. a vendor figure quoted without one tells you nothing about how the detector behaves on real inputs. (more on why that trade-off bites: the false positive rate.)

paraphrasing and humanizers collapse the number

the weakest case for any text detector is text that has been rewritten. a NeurIPS 2023 study showed a paraphrasing model dropped DetectGPT's detection accuracy from 70.3% to 4.6% at a 1% false-positive rate while the meaning stayed intact, and the same trick slipped past watermarking and a commercial classifier. a 2025 follow-up at NeurIPS pushed this harder: an “adversarial paraphrasing” attack cut true-positive rates by an average of about 88% across detectors, and against one popular system by nearly 99%. the “humanizer” tools sold to students run on precisely this gap.

the people most likely to be wrongly flagged are non-native English writers. a 2023 Stanford study published in Patterns ran 91 TOEFL essays by non-native speakers through seven widely used detectors and found, on average, 61.3% were misclassified as AI-generated, against about 5.1% for essays by native-English writers. a fluent yet formulaic style reads as machine-like to these tools, which is why a single flag should never stand in for proof.

why a 99% claim should make you suspicious

regulators have already moved on inflated accuracy claims. in 2025 the U.S. FTC ordered Workado to stop advertising its AI Content Detector as roughly 98% accurate; the FTC's complaint alleged that independent testing put accuracy on general-purpose content at just 53%, because the model had been tuned on academic writing alone. the final order, approved in August 2025, requires competent and reliable evidence behind any accuracy claim. that is the legal reason credible explainers cite third-party studies instead of their own marketing numbers.

tools also vary widely from one another. a 2025 Chicago Booth / Becker Friedman Institute study tested detectors on roughly 2,000 human and 2,000 AI passages across six genres and four frontier models, and found miss rates ranging from near-zero for the strongest tool to somewhere between 10% and 40% for others, with an open-source baseline performing close to random. run the AI text through a humanizer first and miss rates climbed for most of them. a result from one tool on one test set does not generalize to the next.

amige. is built around this evidence rather than against it. every verdict is a probabilistic estimate, not proof, so amige. routes each scan to the detectors strongest for that kind of file, runs a panel of independent detectors built by different teams, and reports a confidence range instead of a bare yes or no. when a model gets named it reads as a best guess (“looks like Midjourney v6”), never a fact, and when the reads conflict the verdict comes back “uncertain.” on text, amige. surfaces the non-native-English false-positive risk instead of hiding it, and recalibrates as new generators land, because the research is clear that yesterday's detector degrades the moment a new model ships. see how the whole pipeline fits together in the machine.